Parallélisation d'un outil de simulation hydraulique numérique

Details

Supervisors

Soares Frazao, Sandra ; Riviere, Etienne

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique

Abstract

enThe SFlow2D program, developed by Pr. S. Soares Frazão, is a hydraulic model to simulate shallow water flows on complex topologies. Hydraulic models used on large-scale, and high resolution problems are computationally expensive. The purpose of this work is therefore to investigate three parallelization methods to reduce execution. These methods are multithreading, the cluster computing and GPU programming. The main challenges are efficient work distribution, information synchronization and centralizing results. Furthermore, work distribution on a cluster requires an additional step : the partitioning of the domain into smaller regions. A good balance in size between regions and minimized borders between regions of the partition allow for reduced data exchange. This partitioning problem has been proven to be NP-hard and has been solved using a heuristic. Two test cases are used in order to validate and quantify the speed-up of our developments. The first is composed of 11,000 cells and represents a simulation of one minute which takes less than 2 minutes to complete. The second case study is based upon the failure of a dam and is composed of more than 60,000 cells. It simulates a flood event of more than 40 hours and takes 54 hours to be generated. Using the cluster has allowed us to reduce the execution time to less than an hour with 48 computational cores. The multithreaded version, takes less than 15 hours to complete (using 8 threads), while the GPU version doesn’t produce concluding results on the Tous test case.